SSMaster’s student in Statistics and Data Science

Early-career designer and developer in training

I am beginning my professional journey, building foundational skills and practical experience to prepare for my first role.

View my workMessage Me

Work samples

I am compiling early projects and practice work that show my developing skills and what I am learning as I prepare for professional work.

  • GI Endoscopy Vision-Language Model (VLM)
    GI Endoscopy Vision-Language Model (VLM)

    Developed a multimodal VLM pipeline combining GI endoscopy image classification, segmentation, and generative language modeling using SigLIP2 and Qwen. Fine-tuned Vision Transformer models on HyperKvasir, achieving ~91% classification accuracy, and developed a Kvasir-SEG segmentation model achieving ~0.90 mean Dice. Implemented and evaluated self-supervised representation learning with DINO, iBOT, and LeJEPA on a 1.1M+ image endoscopy dataset. Built scalable PyTorch training/inference pipelines with Hugging Face, CUDA, cloud GPUs, model checkpointing, evaluation, and interactive VLM deployment.

    SigLIP2QwenVision TransformerHyperKvasirKvasir-SEGDINOiBOTLeJEPAPyTorchHugging FaceCUDAcloud GPUs
  • Machine Learning for Non-Stationary Time Series
    Machine Learning for Non-Stationary Time Series

    Innovated a reweighting framework (MDRIFT) that allows models to adapt to "shifting" data regimes, such as changing lighting or environmental noise, which often degrade standard ML performance. Enhanced an optimization scheme using Difference-of-Convex Algorithms (DCA) to align historical training data with current "target" distributions. Conducted extensive ablation studies comparing TRS and ABS formulations, providing actionable insights into balancing model adaptation with statistical stability.

About me

I am beginning my professional journey, building foundational skills and practical experience to prepare for my first role.

I am Sahibnoor Singh. I have little work experience and have not yet started my career.

Education

Ludwig-Maximilians-Universität München

Statistics and Data Science | Master of Science

Guru Nanak Dev University

Data Science | Bachelor of Science

Skills

Learning mindsetBasic design fundamentalsFundamental development concepts
Back to work

GI Endoscopy Vision-Language Model (VLM)

Developed a multimodal VLM pipeline combining GI endoscopy image classification, segmentation, and generative language modeling using SigLIP2 and Qwen. Fine-tuned Vision Transformer models on HyperKvasir, achieving ~91% classification accuracy, and developed a Kvasir-SEG segmentation model achieving ~0.90 mean Dice. Implemented and evaluated self-supervised representation learning with DINO, iBOT, and LeJEPA on a 1.1M+ image endoscopy dataset. Built scalable PyTorch training/inference pipelines with Hugging Face, CUDA, cloud GPUs, model checkpointing, evaluation, and interactive VLM deployment.

SigLIP2QwenVision TransformerHyperKvasirKvasir-SEGDINOiBOTLeJEPAPyTorchHugging FaceCUDAcloud GPUs
View all work
Back to work

Machine Learning for Non-Stationary Time Series

Innovated a reweighting framework (MDRIFT) that allows models to adapt to "shifting" data regimes, such as changing lighting or environmental noise, which often degrade standard ML performance. Enhanced an optimization scheme using Difference-of-Convex Algorithms (DCA) to align historical training data with current "target" distributions. Conducted extensive ablation studies comparing TRS and ABS formulations, providing actionable insights into balancing model adaptation with statistical stability.

View all work
Back to work

Accent-Aware Speech Recognition

Evaluated the trade-offs of fine-tuning large-scale pretrained models on limited in-domain data, identifying critical failure points where error rates increased by 218%. Constructed and executed a contrastive learning pipeline to investigate how signal-based representations generalize across varied acoustic conditions. Refined training hyperparameters to mitigate the loss of generalizable patterns from pretraining, ensuring better performance in high-noise environments.

View all work
Back to work

Cloud-Native RAG System on Azure

Developed and deployed an end-to-end Retrieval-Augmented Generation (RAG) system that automates the extraction of actionable insights from large-scale technical document repositories. Pioneered a high-performance retrieval pipeline using Azure AI Search for vector-based indexing and FastAPI for low-latency query handling. Enhanced system reliability and portability by containerizing the entire workflow with Docker, ensuring seamless deployment across diverse cloud environments. Refined data quality by implementing advanced text chunking with overlap and embedding generation using state-of-the-art models to ensure grounded, source-cited responses.

Azure AI SearchFastAPIDocker
View all work